Skip to Content
Platform
GuidesOverview

Guides

Step-by-step guides for taking an agent recipe to production and distilling what it learns there back into source you own.

If Introspection is new to you, begin with the Quickstart, then read Core concepts. Come back here for the repeatable guides: each one walks through one outcome and links to the relevant concept, SDK, or CLI reference when you need more detail.

Core workflows

GoalGuideOutcome
Move each recipe change to productionAgent development lifecycleA change proven locally, running as an approved production version.
Put the agent into your productIntegrate into a full-stack appA server or browser integration with identity, reconnect, and production feedback.
Give the agent your source codeWork with repositoriesA runtime that checks out your repositories and can open pull requests against them.
Learn from productionLearn from productionA production signal traced to its source and a hypothesis worth testing.

The continual learning loop in depth

Learn from production opens the loop; three pages continue it, in order:

StageGuideOutcome
Prove the changeEvaluate offlineOne mechanism changed and compared against the baseline at the trace level.
Guard the fixCalibrate judgesA durable risk encoded as a calibrated judge that keeps it fixed.
Distill and shipExperiments & shippingValidated learning merged into the recipe, with a bounded experiment only when offline evidence cannot decide.

Each guide is one outcome end to end. When a step depends on a platform resource — runtimes, judges, experiments, observations — follow its link for the full explanation and return to keep building.

Create projects and connect GitHub in the dashboard. Use the CLI for recipe, runtime, task, and evidence work. Merge recipe changes through Git when they are ready for production.

  • Core concepts: entity definitions and relationships.
  • JS SDK: programmatic contracts used by the examples.
  • CLI: the commands used throughout the guides.
Last updated on